Cause-agnostic bridge damage state identification utilising machine learning

The existing bridge stock, both in the EU and globally, contains several bridges that are reaching the end of their design-life, many of them showing signs of deterioration. Although different deterioration mechanisms are involved in this process, the dominant one is deemed to be the corrosion of tendons. As tendons are not always accessible, their state cannot be interpreted without invasive inspections, thus rendering the structural integrity assessment of bridges with corroded tendons an exceptionally challenging problem. Bridge deterioration is frequently expressed by visible permanent deflections that, however, cannot be interpreted into damage in a straightforward engineering manner due to their cause-agnostic nature. In response to this problem, a new data-driven approach is offered to facilitate a first-order bridge damage state assessment. With the proposed framework one can focus on the end result of the deterioration mechanism that is manifested on the bridge with vertical deck drifts and link these drifts to certain damage levels, which are invaluable in the context of bridge adaptation. An actual balanced cantilever bridge, located in the North-West part of Greece, experiencing large deflections in its balanced cantilever part, is employed to demonstrate the proposed methodology. The methodology involves the use of a finite element model of the bridge to investigate the effect of tendon loss on its structural integrity. Characteristic concrete Young’s modulus values and plausible damage patterns on the prestressing tendons are considered to account for tendon loss and creep effects. A dataset of structural responses is numerically generated, and drift-based fragility functions are obtained. Exploiting the dataset of structural responses developed, the k-Nearest Neighbours (k-NN) machine learning algorithm is then deployed to rapidly identify the damage state of bridges that would otherwise require thorough inspections and testing. The input required for the identification of the bridge damage state is only the observed deflected shape and the measured concrete Young’s modulus. Thus, the significance of the proposed methodology is its ability to interpret vertical deck deflections, associated with tendon loss, into a bridge damage level, invaluable to decisions toward bridge retrofitting and adaptation. ● Novel methodology for bridge damage state assessment. ● Monitoring data processed through machine learning to link deflections to tendon loss scenarios. ● New deck drift-based fragility functions for cause-agnostic bridge damage state identification. ● Cost and time efficient methodology requiring very limited physical interaction with the bridge. ● Transferable methodology for structural damage state-informed action planning.